diff --git a/.sync_state b/.sync_state
index 0488b81..bffd7ba 100644
--- a/.sync_state
+++ b/.sync_state
@@ -1,4 +1,4 @@
 {
-  "last_synced_sha": "04bf337b7d4063a8f0a7b2c087a823e496e8c626",
-  "last_sync_time": "2025-11-25T05:50:26.440746"
+  "last_synced_sha": "110dad99af8ae1062af4fa99fc7559104d793680",
+  "last_sync_time": "2025-11-25T05:55:10.370139"
 }
\ No newline at end of file
diff --git a/docs/api/apifuture.md b/docs/api/apifuture.md
index 871dfa2..f6059b3 100644
--- a/docs/api/apifuture.md
+++ b/docs/api/apifuture.md
@@ -25,7 +25,7 @@ result = api_future.result()  # Blocks until complete
 ```
 
 Args:
-    T: The type of the result value
+- `T`: The type of the result value
 
 Example:
 ```python
@@ -48,10 +48,10 @@ async def result_async(timeout: float | None = None) -> T
 Get the result asynchronously with optional timeout.
 
 Args:
-    timeout: Maximum time to wait in seconds. None means wait indefinitely.
+- `timeout`: Maximum time to wait in seconds. None means wait indefinitely.
 
 Returns:
-    The result value of type T
+- The result value of type `T`
 
 Raises:
     TimeoutError: If timeout is exceeded
@@ -66,10 +66,10 @@ def result(timeout: float | None = None) -> T
 Get the result synchronously with optional timeout.
 
 Args:
-    timeout: Maximum time to wait in seconds. None means wait indefinitely.
+- `timeout`: Maximum time to wait in seconds. None means wait indefinitely.
 
 Returns:
-    The result value of type T
+- The result value of type `T`
 
 Raises:
     TimeoutError: If timeout is exceeded
@@ -87,7 +87,7 @@ standard Future to be used in async contexts. It's commonly returned by
 Tinker API methods to provide both sync and async access patterns.
 
 Args:
-    future: A concurrent.futures.Future to wrap
+- `future`: A concurrent.futures.Future to wrap
 
 Example:
 ```python
@@ -111,10 +111,10 @@ def result(timeout: float | None = None) -> T
 Get the result synchronously with optional timeout.
 
 Args:
-    timeout: Maximum time to wait in seconds. None means wait indefinitely.
+- `timeout`: Maximum time to wait in seconds. None means wait indefinitely.
 
 Returns:
-    The result value of type T
+- The result value of type `T`
 
 Raises:
     TimeoutError: If timeout is exceeded
@@ -143,7 +143,7 @@ def future() -> ConcurrentFuture[T]
 Get the underlying concurrent.futures.Future.
 
 Returns:
-    The wrapped ConcurrentFuture object
+- The wrapped `ConcurrentFuture` object
 
 Example:
 ```python
diff --git a/docs/api/restclient.md b/docs/api/restclient.md
index d50723c..5bc7509 100644
--- a/docs/api/restclient.md
+++ b/docs/api/restclient.md
@@ -24,7 +24,7 @@ Key methods:
 - unpublish_checkpoint_from_tinker_path() - unpublish a checkpoint to make it private
 
 Args:
-    holder: Internal client managing HTTP connections and async operations
+- `holder`: Internal client managing HTTP connections and async operations
 
 Example:
 ```python
@@ -49,10 +49,10 @@ def get_training_run(
 Get training run info.
 
 Args:
-    training_run_id: The training run ID to get information for
+- `training_run_id`: The training run ID to get information for
 
 Returns:
-    A Future containing the training run information
+- A `Future` containing the training run information
 
 Example:
 ```python
@@ -83,10 +83,10 @@ def get_training_run_by_tinker_path(
 Get training run info.
 
 Args:
-    tinker_path: The tinker path to the checkpoint
+- `tinker_path`: The tinker path to the checkpoint
 
 Returns:
-    A Future containing the training run information
+- A `Future` containing the training run information
 
 Example:
 ```python
@@ -116,10 +116,10 @@ def get_weights_info_by_tinker_path(
 Get checkpoint information from a tinker path.
 
 Args:
-    tinker_path: The tinker path to the checkpoint
+- `tinker_path`: The tinker path to the checkpoint
 
 Returns:
-    An APIFuture containing the checkpoint information. The future is awaitable.
+- An `APIFuture` containing the checkpoint information. The future is awaitable.
 
 Example:
 ```python
@@ -141,11 +141,11 @@ def list_training_runs(
 List training runs with pagination support.
 
 Args:
-    limit: Maximum number of training runs to return (default 20)
-    offset: Offset for pagination (default 0)
+- `limit`: Maximum number of training runs to return (default 20)
+- `offset`: Offset for pagination (default 0)
 
 Returns:
-    A Future containing the TrainingRunsResponse with training runs and cursor info
+- A `Future` containing the `TrainingRunsResponse` with training runs and cursor info
 
 Example:
 ```python
@@ -181,10 +181,10 @@ def list_checkpoints(
 List available checkpoints (both training and sampler).
 
 Args:
-    training_run_id: The training run ID to list checkpoints for
+- `training_run_id`: The training run ID to list checkpoints for
 
 Returns:
-    A Future containing the CheckpointsListResponse with available checkpoints
+- A `Future` containing the `CheckpointsListResponse` with available checkpoints
 
 Example:
 ```python
@@ -220,11 +220,11 @@ def get_checkpoint_archive_url(
 Get signed URL to download checkpoint archive.
 
 Args:
-    training_run_id: The training run ID to download weights for
-    checkpoint_id: The checkpoint ID to download
+- `training_run_id`: The training run ID to download weights for
+- `checkpoint_id`: The checkpoint ID to download
 
 Returns:
-    A Future containing the CheckpointArchiveUrlResponse with signed URL and expiration
+- A `Future` containing the `CheckpointArchiveUrlResponse` with signed URL and expiration
 
 Example:
 ```python
@@ -300,10 +300,10 @@ def get_checkpoint_archive_url_from_tinker_path(
 Get signed URL to download checkpoint archive.
 
 Args:
-    tinker_path: The tinker path to the checkpoint
+- `tinker_path`: The tinker path to the checkpoint
 
 Returns:
-    A Future containing the CheckpointArchiveUrlResponse with signed URL and expiration
+- A `Future` containing the `CheckpointArchiveUrlResponse` with signed URL and expiration
 
 #### `get_checkpoint_archive_url_from_tinker_path_async`
 
@@ -330,10 +330,10 @@ Only the exact owner of the training run can publish checkpoints.
 Published checkpoints can be unpublished using the unpublish_checkpoint_from_tinker_path method.
 
 Args:
-    tinker_path: The tinker path to the checkpoint (e.g., "tinker://run-id/weights/0001")
+- `tinker_path`: The tinker path to the checkpoint (e.g., "tinker://run-id/weights/0001")
 
 Returns:
-    A Future that completes when the checkpoint is published
+- A `Future` that completes when the checkpoint is published
 
 Raises:
     HTTPException: 400 if checkpoint identifier is invalid
@@ -372,10 +372,10 @@ Only the exact owner of the training run can unpublish checkpoints.
 This reverses the effect of publishing a checkpoint.
 
 Args:
-    tinker_path: The tinker path to the checkpoint (e.g., "tinker://run-id/weights/0001")
+- `tinker_path`: The tinker path to the checkpoint (e.g., "tinker://run-id/weights/0001")
 
 Returns:
-    A Future that completes when the checkpoint is unpublished
+- A `Future` that completes when the checkpoint is unpublished
 
 Raises:
     HTTPException: 400 if checkpoint identifier is invalid
@@ -417,11 +417,11 @@ sorted by time (newest first). It supports pagination for efficiently handling l
 numbers of checkpoints.
 
 Args:
-    limit: Maximum number of checkpoints to return (default 100)
-    offset: Offset for pagination (default 0)
+- `limit`: Maximum number of checkpoints to return (default 100)
+- `offset`: Offset for pagination (default 0)
 
 Returns:
-    A Future containing the CheckpointsListResponse with checkpoints and cursor info
+- A `Future` containing the `CheckpointsListResponse` with checkpoints and cursor info
 
 Example:
 ```python
@@ -458,10 +458,10 @@ def get_session(session_id: str) -> ConcurrentFuture[types.GetSessionResponse]
 Get session information including all training runs and samplers.
 
 Args:
-    session_id: The session ID to get information for
+- `session_id`: The session ID to get information for
 
 Returns:
-    A Future containing the GetSessionResponse with training_run_ids and sampler_ids
+- A `Future` containing the `GetSessionResponse` with training_run_ids and sampler_ids
 
 Example:
 ```python
@@ -493,11 +493,11 @@ def list_sessions(
 List sessions with pagination support.
 
 Args:
-    limit: Maximum number of sessions to return (default 20)
-    offset: Offset for pagination (default 0)
+- `limit`: Maximum number of sessions to return (default 20)
+- `offset`: Offset for pagination (default 0)
 
 Returns:
-    A Future containing the ListSessionsResponse with list of session IDs
+- A `Future` containing the `ListSessionsResponse` with list of session IDs
 
 Example:
 ```python
@@ -528,10 +528,10 @@ def get_sampler(sampler_id: str) -> APIFuture[types.GetSamplerResponse]
 Get sampler information.
 
 Args:
-    sampler_id: The sampler ID (sampling_session_id) to get information for
+- `sampler_id`: The sampler ID (sampling_session_id) to get information for
 
 Returns:
-    An APIFuture containing the GetSamplerResponse with sampler details
+- An `APIFuture` containing the `GetSamplerResponse` with sampler details
 
 Example:
 ```python
diff --git a/docs/api/samplingclient.md b/docs/api/samplingclient.md
index 95f95a3..e7183ec 100644
--- a/docs/api/samplingclient.md
+++ b/docs/api/samplingclient.md
@@ -18,10 +18,10 @@ Key methods:
 - compute_logprobs() - get log probabilities for prompt tokens
 
 Args:
-    holder: Internal client managing HTTP connections and async operations
-    model_path: Path to saved model weights (starts with 'tinker://')
-    base_model: Name of base model to use for inference
-    retry_config: Configuration for retrying failed requests
+- `holder`: Internal client managing HTTP connections and async operations
+- `model_path`: Path to saved model weights (starts with 'tinker://')
+- `base_model`: Name of base model to use for inference
+- `retry_config`: Configuration for retrying failed requests
 
 Example:
 ```python
@@ -48,14 +48,14 @@ def sample(
 Generate text completions from the model.
 
 Args:
-    prompt: The input tokens as ModelInput
-    num_samples: Number of independent samples to generate
-    sampling_params: Parameters controlling generation (temperature, max_tokens, etc.)
-    include_prompt_logprobs: Whether to include log probabilities for prompt tokens
-    topk_prompt_logprobs: Number of top token log probabilities to return per position
+- `prompt`: The input tokens as ModelInput
+- `num_samples`: Number of independent samples to generate
+- `sampling_params`: Parameters controlling generation (temperature, max_tokens, etc.)
+- `include_prompt_logprobs`: Whether to include log probabilities for prompt tokens
+- `topk_prompt_logprobs`: Number of top token log probabilities to return per position
 
 Returns:
-    A Future containing the SampleResponse with generated text
+- A `Future` containing the `SampleResponse` with generated text
 
 Example:
 ```python
@@ -90,10 +90,10 @@ def compute_logprobs(
 Compute log probabilities for prompt tokens.
 
 Args:
-    prompt: The input tokens as ModelInput
+- `prompt`: The input tokens as ModelInput
 
 Returns:
-    A Future containing a list of log probabilities for each token in the prompt.
+- A `Future` containing a list of log probabilities for each token in the prompt.
     None values indicate tokens where log probabilities couldn't be computed.
 
 Example:
diff --git a/docs/api/serviceclient.md b/docs/api/serviceclient.md
index cd0a9e4..da814d8 100644
--- a/docs/api/serviceclient.md
+++ b/docs/api/serviceclient.md
@@ -20,14 +20,17 @@ Args:
 
 Example:
 ```python
+# Near instant
 client = ServiceClient()
-# ^^^ near-instant
+
+# Takes a moment as we initialize the model and assign resources
 training_client = client.create_lora_training_client(base_model="Qwen/Qwen3-8B")
-# ^^^ takes a moment as we initialize the model and assign resources
+
+# Near-instant
 sampling_client = client.create_sampling_client(base_model="Qwen/Qwen3-8B")
-# ^^^ near-instant
+
+# Near-instant
 rest_client = client.create_rest_client()
-# ^^^ near-instant
 ```
 
 #### `get_server_capabilities`
@@ -41,7 +44,7 @@ def get_server_capabilities() -> types.GetServerCapabilitiesResponse
 Query the server's supported features and capabilities.
 
 Returns:
-    GetServerCapabilitiesResponse with available models, features, and limits
+- `GetServerCapabilitiesResponse` with available models, features, and limits
 
 Example:
 ```python
@@ -78,16 +81,16 @@ def create_lora_training_client(
 Create a TrainingClient for LoRA fine-tuning.
 
 Args:
-    base_model: Name of the base model to fine-tune (e.g., "Qwen/Qwen2.5-7B")
-    rank: LoRA rank controlling the size of adaptation matrices (default 32)
-    seed: Random seed for initialization. None means random seed.
-    train_mlp: Whether to train MLP layers (default True)
-    train_attn: Whether to train attention layers (default True)
-    train_unembed: Whether to train unembedding layers (default True)
-    user_metadata: Optional metadata to attach to the training run
+- `base_model`: Name of the base model to fine-tune (e.g., "Qwen/Qwen2.5-7B")
+- `rank`: LoRA rank controlling the size of adaptation matrices (default 32)
+- `seed`: Random seed for initialization. None means random seed.
+- `train_mlp`: Whether to train MLP layers (default True)
+- `train_attn`: Whether to train attention layers (default True)
+- `train_unembed`: Whether to train unembedding layers (default True)
+- `user_metadata`: Optional metadata to attach to the training run
 
 Returns:
-    TrainingClient configured for LoRA training
+- `TrainingClient` configured for LoRA training
 
 Example:
 ```python
@@ -129,11 +132,11 @@ def create_training_client_from_state(
 Create a TrainingClient from saved model weights.
 
 Args:
-    path: Tinker path to saved weights (e.g., "tinker://run-id/weights/checkpoint-001")
-    user_metadata: Optional metadata to attach to the new training run
+- `path`: Tinker path to saved weights (e.g., "tinker://run-id/weights/checkpoint-001")
+- `user_metadata`: Optional metadata to attach to the new training run
 
 Returns:
-    TrainingClient loaded with the specified weights
+- `TrainingClient` loaded with the specified weights
 
 Example:
 ```python
@@ -168,12 +171,12 @@ def create_sampling_client(
 Create a SamplingClient for text generation.
 
 Args:
-    model_path: Path to saved model weights (e.g., "tinker://run-id/weights/checkpoint-001")
-    base_model: Name of base model to use (e.g., "Qwen/Qwen2.5-7B")
-    retry_config: Optional configuration for retrying failed requests
+- `model_path`: Path to saved model weights (e.g., "tinker://run-id/weights/checkpoint-001")
+- `base_model`: Name of base model to use (e.g., "Qwen/Qwen2.5-7B")
+- `retry_config`: Optional configuration for retrying failed requests
 
 Returns:
-    SamplingClient configured for text generation
+- `SamplingClient` configured for text generation
 
 Raises:
     ValueError: If neither model_path nor base_model is provided
@@ -216,7 +219,7 @@ The RestClient provides access to various REST endpoints for querying
 model information, checkpoints, sessions, and managing checkpoint visibility.
 
 Returns:
-    RestClient for accessing REST API endpoints
+- `RestClient` for accessing REST API endpoints
 
 Example:
 ```python
diff --git a/docs/api/trainingclient.md b/docs/api/trainingclient.md
index 74a0b7d..1f5f420 100644
--- a/docs/api/trainingclient.md
+++ b/docs/api/trainingclient.md
@@ -18,8 +18,8 @@ Key methods:
 - save_weights_and_get_sampling_client() - export trained model for inference
 
 Args:
-    holder: Internal client managing HTTP connections and async operations
-    model_id: Unique identifier for the model to train. Required for training operations.
+- `holder`: Internal client managing HTTP connections and async operations
+- `model_id`: Unique identifier for the model to train. Required for training operations.
 
 Example:
 ```python
@@ -45,12 +45,12 @@ def forward(
 Compute forward pass without gradients.
 
 Args:
-    data: List of training data samples
-    loss_fn: Loss function type (e.g., "cross_entropy")
-    loss_fn_config: Optional configuration for the loss function
+- `data`: List of training data samples
+- `loss_fn`: Loss function type (e.g., "cross_entropy")
+- `loss_fn_config`: Optional configuration for the loss function
 
 Returns:
-    APIFuture containing the forward pass outputs and loss
+- `APIFuture` containing the forward pass outputs and loss
 
 Example:
 ```python
@@ -89,12 +89,12 @@ def forward_backward(
 Compute forward pass and backward pass to calculate gradients.
 
 Args:
-    data: List of training data samples
-    loss_fn: Loss function type (e.g., "cross_entropy")
-    loss_fn_config: Optional configuration for the loss function
+- `data`: List of training data samples
+- `loss_fn`: Loss function type (e.g., "cross_entropy")
+- `loss_fn_config`: Optional configuration for the loss function
 
 Returns:
-    APIFuture containing the forward/backward outputs, loss, and gradients
+- `APIFuture` containing the forward/backward outputs, loss, and gradients
 
 Example:
 ```python
@@ -143,11 +143,11 @@ Allows you to define custom loss functions that operate on log probabilities.
 The custom function receives logprobs and computes loss and gradients.
 
 Args:
-    data: List of training data samples
-    loss_fn: Custom loss function that takes (data, logprobs) and returns (loss, metrics)
+- `data`: List of training data samples
+- `loss_fn`: Custom loss function that takes (data, logprobs) and returns (loss, metrics)
 
 Returns:
-    APIFuture containing the forward/backward outputs with custom loss
+- `APIFuture` containing the forward/backward outputs with custom loss
 
 Example:
 ```python
@@ -185,10 +185,10 @@ def optim_step(
 Update model parameters using Adam optimizer.
 
 Args:
-    adam_params: Adam optimizer parameters (learning_rate, betas, eps, weight_decay)
+- `adam_params`: Adam optimizer parameters (learning_rate, betas, eps, weight_decay)
 
 Returns:
-    APIFuture containing optimizer step response
+- `APIFuture` containing optimizer step response
 
 Example:
 ```python
@@ -227,10 +227,10 @@ def save_state(name: str) -> APIFuture[types.SaveWeightsResponse]
 Save model weights to persistent storage.
 
 Args:
-    name: Name for the saved checkpoint
+- `name`: Name for the saved checkpoint
 
 Returns:
-    APIFuture containing the save response with checkpoint path
+- `APIFuture` containing the save response with checkpoint path
 
 Example:
 ```python
@@ -258,10 +258,10 @@ def load_state(path: str) -> APIFuture[types.LoadWeightsResponse]
 Load model weights from a saved checkpoint.
 
 Args:
-    path: Tinker path to saved weights (e.g., "tinker://run-id/weights/checkpoint-001")
+- `path`: Tinker path to saved weights (e.g., "tinker://run-id/weights/checkpoint-001")
 
 Returns:
-    APIFuture containing the load response
+- `APIFuture` containing the load response
 
 Example:
 ```python
@@ -290,10 +290,10 @@ def load_state_with_optimizer(
 Load model weights and optimizer state from a checkpoint.
 
 Args:
-    path: Tinker path to saved weights (e.g., "tinker://run-id/weights/checkpoint-001")
+- `path`: Tinker path to saved weights (e.g., "tinker://run-id/weights/checkpoint-001")
 
 Returns:
-    APIFuture containing the load response
+- `APIFuture` containing the load response
 
 Example:
 ```python
@@ -325,10 +325,10 @@ def save_weights_for_sampler(
 Save model weights for use with a SamplingClient.
 
 Args:
-    name: Name for the saved sampler weights
+- `name`: Name for the saved sampler weights
 
 Returns:
-    APIFuture containing the save response with sampler path
+- `APIFuture` containing the save response with sampler path
 
 Example:
 ```python
@@ -363,7 +363,7 @@ def get_info() -> types.GetInfoResponse
 Get information about the current model.
 
 Returns:
-    GetInfoResponse with model configuration and metadata
+- `GetInfoResponse` with model configuration and metadata
 
 Example:
 ```python
@@ -393,7 +393,7 @@ def get_tokenizer() -> PreTrainedTokenizer
 Get the tokenizer for the current model.
 
 Returns:
-    PreTrainedTokenizer compatible with the model
+- `PreTrainedTokenizer` compatible with the model
 
 Example:
 ```python
@@ -414,11 +414,11 @@ def create_sampling_client(
 Create a SamplingClient from saved weights.
 
 Args:
-    model_path: Tinker path to saved weights
-    retry_config: Optional configuration for retrying failed requests
+- `model_path`: Tinker path to saved weights
+- `retry_config`: Optional configuration for retrying failed requests
 
 Returns:
-    SamplingClient configured with the specified weights
+- `SamplingClient` configured with the specified weights
 
 Example:
 ```python
@@ -451,11 +451,11 @@ def save_weights_and_get_sampling_client(
 Save current weights and create a SamplingClient for inference.
 
 Args:
-    name: Optional name for the saved weights (currently ignored for ephemeral saves)
-    retry_config: Optional configuration for retrying failed requests
+- `name`: Optional name for the saved weights (currently ignored for ephemeral saves)
+- `retry_config`: Optional configuration for retrying failed requests
 
 Returns:
-    SamplingClient configured with the current model weights
+- `SamplingClient` configured with the current model weights
 
 Example:
 ```python